Diplomarbeit: Learning Policies for Reliable Mobile Robot Localization
Abstract
Cameras are a useful sensor for mobile robot localization because they are relatively cheap, compact, and lightweight. This makes them attractive for robots with payload limitations, such as humanoid robots or small unmanned aerial vehicles. However, fast movements typically reduce the ability to use a vision-based localization due to motion blur.
In this thesis, we present a reinforcement learning approach for a robot learning a vision-based navigation policy. The learned policy minimizes the time to reach the destination and implicitly takes the impact of motion blur on landmark observations and thus on localization into account. Extensive simulated and real-world experiments show that our learned policy significantly outperforms any policy of moving with a constant velocity and is generally applicable to different environments. We experimentally determined the most relevant state features for the learning task.
Additionally, we present a method for compressing the learned policy with a clustering approach. While the size of the policy representation is drastically reduced, our experiments show that there is no loss of performance. This is especially valuable for memory-constrained systems.
Complete thesis (95 pages, 4.4 MB)